Tensorflow implementation of LSTM based sentiment analysis with attention layer

Bag of Words vs. Bag of Popcorn as an example

This is an extension of the previous project (lstm_word2vec). Now we have a tensorflow implementation of LSTM for the purpose of sentiment analysis, with the option of adding an attention layer.

In this project, I reused the I/O function and data prep functions from lstm_word2vec.

The major improvements are:

1. Migration of code from keras to TF.
2. Added an optional attention layer plus visualization.
3. Added a batch generator class that can be easily used for different datasets.

The main_lstm.ipynb show cases how to leverage existing functions to train your own LSTM + attention layer model.

Addition of the attention layer appears to marginally improve the results of the testing set. But further testing is needed to establish the significance of the improvement...